Background Single-cell RNA sequencing (scRNA-seq) enables characterization of cellular heterogeneity, but accurate and stable clustering remains challenging due to high dimensionality, sparsity, and technical noise. Many existing methods insufficiently exploit intrinsic cell–cell relationships and latent cell-type signals, leading to suboptimal representations and unstable assignments. The objective of this study is to develop a self-supervised graph contrastive framework for scRNA-seq clustering that jointly leverages augmented views, intrinsic cell–cell relationships, and pseudo-label-guided graph refinement to improve clustering robustness and biological interpretability. Methods We propose Self-Supervised Contrastive Graph Learning (SSGL), which refers specifically to the proposed scRNA-seq clustering framework and not to a generic or pre-existing method. SSGL generates two augmented views of each cell via dual random gene masking, then learns representations with a momentum-encoder architecture. A refined cell–cell graph is constructed by intersecting a k-nearest-neighbor similarity graph with pseudo-label consistency, and a graph-aware contrastive objective enforces agreement between augmented views while preserving local neighborhood structure. Results We evaluated SSGL on eight public scRNA-seq benchmarks spanning different tissues and sequencing platforms, and compared it with representative clustering baselines using NMI and ARI. Across all datasets, SSGL achieved the best overall clustering performance, with average NMI = 0.876 and average ARI = 0.926, outperforming competing approaches and showing more consistent results across heterogeneous datasets. Relative to AttentionAE-SC, the strongest overall baseline based on the averaged results, SSGL improved the average NMI and ARI by approximately 4.4% and 6.7%, respectively. Ablation analyses demonstrated that incorporating the self-supervised refined graph improves clustering beyond conventional contrastive objectives. Visualization and marker-gene analyses further supported that SSGL yields compact, well-separated clusters and recovers biologically coherent cell groups, including rare populations. Conclusions SSGL improves scRNA-seq clustering accuracy and stability by jointly leveraging augmented views, intrinsic cell–cell relationships, and pseudo-label-guided graph refinement within a graph contrastive learning framework. This provides robust representations that support reliable cell-type discovery and downstream biological interpretation.

Self-supervised graph contrastive learning for scRNA-seq clustering

期刊论文
出版模式：

开放获取
发表日期：

2026年6月23日

查看全文

Tong Wu

摘要

Background

Single-cell RNA sequencing (scRNA-seq) enables characterization of cellular heterogeneity, but accurate and stable clustering remains challenging due to high dimensionality, sparsity, and technical noise. Many existing methods insufficiently exploit intrinsic cell–cell relationships and latent cell-type signals, leading to suboptimal representations and unstable assignments. The objective of this study is to develop a self-supervised graph contrastive framework for scRNA-seq clustering that jointly leverages augmented views, intrinsic cell–cell relationships, and pseudo-label-guided graph refinement to improve clustering robustness and biological interpretability.

Methods

We propose Self-Supervised Contrastive Graph Learning (SSGL), which refers specifically to the proposed scRNA-seq clustering framework and not to a generic or pre-existing method. SSGL generates two augmented views of each cell via dual random gene masking, then learns representations with a momentum-encoder architecture. A refined cell–cell graph is constructed by intersecting a k-nearest-neighbor similarity graph with pseudo-label consistency, and a graph-aware contrastive objective enforces agreement between augmented views while preserving local neighborhood structure.

Results

We evaluated SSGL on eight public scRNA-seq benchmarks spanning different tissues and sequencing platforms, and compared it with representative clustering baselines using NMI and ARI. Across all datasets, SSGL achieved the best overall clustering performance, with average NMI = 0.876 and average ARI = 0.926, outperforming competing approaches and showing more consistent results across heterogeneous datasets. Relative to AttentionAE-SC, the strongest overall baseline based on the averaged results, SSGL improved the average NMI and ARI by approximately 4.4% and 6.7%, respectively. Ablation analyses demonstrated that incorporating the self-supervised refined graph improves clustering beyond conventional contrastive objectives. Visualization and marker-gene analyses further supported that SSGL yields compact, well-separated clusters and recovers biologically coherent cell groups, including rare populations.

Conclusions

SSGL improves scRNA-seq clustering accuracy and stability by jointly leveraging augmented views, intrinsic cell–cell relationships, and pseudo-label-guided graph refinement within a graph contrastive learning framework. This provides robust representations that support reliable cell-type discovery and downstream biological interpretation.